Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests

Anthropic disclosed in a safety report that its internal filtering system designed to block biological and chemical weapons inquiries remained inactive for nearly a year. During this period, approximately 50,000 external feedback contractors processed roughly 133 million unfiltered interactions with the company's AI models.

The failure represents a significant gap in Anthropic's safety infrastructure. The company relies on these filters as a core control mechanism to prevent its models from providing information that could facilitate weapons development or misuse. With the system offline, all queries related to bio-weapons and chemical weapons risks bypassed the intended safeguards.

Anthropic discovered the issue and has since reactivated the filtering system. The company did not disclose whether any of the 133 million interactions resulted in problematic outputs or posed genuine security risks. It also remains unclear what caused the system to go offline or why the outage persisted undetected for such an extended period.

The disclosure raises questions about Anthropic's monitoring procedures and internal oversight. A safety control this fundamental should trigger alerts when disabled. The extended downtime suggests either insufficient redundancy in the detection process or gaps in how the company tracks its safety infrastructure status.

Anthropic has positioned itself as a leader in AI safety, publishing extensive research on alignment and risk mitigation. This incident contradicts that messaging. Even with the best theoretical safety practices, execution failures can undermine the entire approach.

The company did not specify whether it informed regulators or customers about the outage. As AI models become more powerful and more widely deployed, transparency about safety failures becomes increasingly important. Users of Anthropic's API and Claude interface during the outage period have no way to verify whether their interactions were secure.

This incident underscores a broader challenge in AI safety: maintaining safety systems requires constant operational diligence, not just good intentions. The industry lacks universal standards for monitoring critical safety controls, and companies often